Magnetic Anomaly Detection Using Multifeature Fusion-Based Neural Network

Magnetic anomaly detection (MAD) is widely applied in the fields of resource exploration, hidden target detection, and explosive ordnance disposal. Traditional methods, such as orthonormal basis functions (OBFs), are proposed to extract anomaly signals from ambient noises and device noises. Due to the weakness of the signal, the detection probability has always been limited by a low signal-to-noise ratio (SNR). To surmount the limitation, a full connected neural network (FCN) with OBF features is trained to do the detection. Nonetheless, its effect is not reliable enough under a low SNR, and it is sensitive to the orientations. This letter introduces a multifeature fusion-based neural network with three subclassifiers to conduct MAD. The first subclassifier uses the time-frequency feature, the second uses the statistical feature, and the last concentrates on the magnetic moment feature. The outputs of the subclassifiers are analyzed synthetically by weighted voting, and the optimized weights are picked on the basis of individual performance. The real noise is recorded by experiments to test the performance of our network. The result indicates that a multifeature-based neural network shows a higher detection probability than the ordinary FCN by 5% medially. At very low SNR, the multifeature-based neural network can achieve a detection probability 13% higher than FCN. Sensitivity to orientations is also improved by the multifeature-based neural network.